AI4S is not short of stories, but what it lacks is how to create real business value.

As AI for Science gradually moves from concept to implementation, "industrial capabilities" have become an inescapable yardstick for the entire industry. All scientific problems ultimately need to be verified in the real world - they need to enter fermentation tanks, factories, supply chains, and customers' production processes.

But the core question is, what exactly does the commercialization path of AI4S look like? From the first laboratory result to the first continuously repurchased order, what real obstacles need to be overcome, and what trade - offs need to be made?

Recently, at an event co - hosted by Future Tech of the World Artificial Intelligence Conference, Linear Capital, and the Institute of Industry - Academia - Research of Shanghai Jiao Tong University, Zeng Yingzhe, a partner at Linear Capital, along with Su Rui, the founder of Yiru Bio, and Zhang Luyang, the COO of Deep Principle, shared how they use AI to compress trial - and - error costs, accumulate data flywheels, and find commercial footholds in the AI4S field in real industrial scenarios.

For a long time in the past, when people discussed AI4S, they were more focused on whether it could "discover new things." For example, AlphaFold predicts protein structures, large models screen new materials, and AI finds a previously unnoticed molecular combination.

But for those truly engaged in the industry, "discovery" is not always the most difficult part.

The real difficulties also include: after a molecule is discovered, how can it be mass - produced stably in a 20 - ton fermentation tank? How can a laboratory result overcome the constraints in processes, equipment, and real - world scenarios and ultimately become a product that customers are willing to pay for continuously?

In the field of biomanufacturing, this complexity is particularly evident. Today, AI can make a large number of predictions at the molecular level. However, when a strain truly enters the industrial fermentation system, the variables will expand rapidly: pressure differentials, dissolved oxygen, flow - addition speed, agitator design, and even some microscopic changes that are difficult to fully explain can all lead to the failure of the entire fermentation process.

In the past, these problems highly relied on the "intuition" of experienced engineers and long - term trial and error. Today, AI4S is starting to change this situation.

When the test of AI for Science expands from "model capabilities" to "industrial capabilities", all those scientific problems ultimately need to be verified in the real world - they need to enter fermentation tanks, factories, supply chains, and customers' production processes. A core question is, what exactly does the commercialization path of AI4S look like? From the first laboratory result to the first continuously repurchased order, what real obstacles need to be overcome, and what trade - offs need to be made?

And in this process, China's complete manufacturing system, industrial scenarios, and engineering capabilities are becoming an increasingly important and difficult - to - replicate advantage in the global AI4S competition. The following are some highlights of this dialogue:

From molecules to fermentation tanks, what impossible problems is AI solving?

Su Rui @ Yiru Bio

Yiru Bio was founded in 2021. It is a biomanufacturing enterprise that mainly uses microorganisms to produce elastomer materials. When we first started our business, there were not as many AI - related concepts. But since we are engaged in biomanufacturing, after the emergence of ChatGPT, we have gradually explored how to use AI to empower the entire process in the traditional industry.

Today, I mainly want to share our understanding of AI and some opportunities and explorations for overtaking on the curve in the field of production and manufacturing, especially in the process of scaling up from the laboratory to the factory scale.

Fermentation scale - up is a well - recognized extremely difficult problem in synthetic biology. Tools like AlphaFold mainly solve problems at the molecular level. Since DNA itself is a kind of language, after it is translated into proteins, the subsequent process still belongs to the language system. This is actually relatively easy in the industry. However, in the field of biomanufacturing, the situation is completely different.

In biomanufacturing, the most difficult thing has never been to discover a valuable product. For example, among the product selections we are currently focusing on, there may already be more than 60 kinds in our reserve, covering various materials and molecules, all of which are worth pursuing. But the most core pain point lies in how to truly scale them up to an industrial level.

Let me give a simple example. It took us nearly four years to go from the initial bio - elastic cohesive molecules to the mass production in a 20 - ton fermentation tank. During this period, we abandoned countless solutions.

The biggest problem here is that biology itself is a highly complex "mysterious problem." We often joke that today, maybe because we "didn't worship a certain god," the fermentation tank fails. Because there are too many variables, you can never predict what kind of chain reaction will occur when organisms are at an extremely high density of 10^12 to 10^13.

Using existing technical means, such as controlling pH value, dissolved oxygen, or fine - tuned flow - addition, can solve some problems. However, when the scale is expanded to fermentation tanks larger than 20 tons, higher - dimensional challenges will arise.

A classic example is that yeast may not be able to withstand the pressure differential of the liquid level, or the design of the agitator may cause the linear velocity to be too high, directly "killing" the cells or bacteria.

In the past, to solve these problems, it purely relied on the experience and intuition of traditional engineers. They were like "veteran masters" controlling the industrial details inside, which cannot be simulated by existing ordinary models.

Therefore, we developed a system for specific strains that can simulate the fermentation performance from a 5 - liter R & D scale to the maximum 20 - ton industrial scale. We use various algorithms for composite modeling. Since the Agent in AI has strong autonomy, it can write hundreds of mathematical models and combine them. Through this modeling, the system will ultimately form an approximately predictable pattern.

So we believe that using AI to discover and accelerate the laws of fermentation scale - up is very meaningful.

Let me give another simple example. In the past, to scale up to a 20 - ton scale, we might have to conduct 7 to 8 consecutive trial - and - error batches to figure out the pattern. In industrial - scale fermentation, the trial - and - error cost of one batch of experiments is very high.

But if we introduce AI, we can conduct simulations in a 200 - liter micro - system. One experiment may only cost a few thousand dollars, but it can simulate the process of a large - scale system, thus accurately guiding our future actual fermentation production. This is the first part, that is, to improve R & D efficiency through real cost - reduction and efficiency - improvement.

The second logic is that it will form a new business closed - loop for biotech companies. We have always been thinking about where the truly valuable assets of future AI companies lie. We believe it lies in the data of vertical industrial scenarios.

Now, every time we complete a fermentation, the built - in adaptive module of the system will input the newly generated data, allowing the model to continuously learn. As the data accumulates, the model's predictions will become more and more stable. In the highly complex field of biology, the support of AI in such specific scenarios has extremely high commercial value and can truly empower many application scenarios.

Zhang Luyang @ Deep Principle

Deep Principle was founded in 2024. It mainly uses Deep Learning combined with first - principles calculations to unlock new materials and scientific discoveries. Before becoming the COO of Deep Principle, I also worked at Horizon Robotics.

I majored in semiconductors. In this field, there is Moore's Law. Moore's Law means that the growth rate of computing power is very fast. However, in contrast, the R & D efficiency of drug and material discovery is getting lower and lower.

We then found that there is actually a good opportunity here: we can use exponentially growing computing power to make up for an industry with decreasing R & D efficiency.

So from the very beginning, we targeted this direction. For materials, there are 10^60 different possible arrangements in the entire chemical space, resulting in very sparse data. It is unrealistic to directly use a large model to solve all problems. Therefore, we firmly believe that what we do must be able to be implemented and solve specific problems.

Based on this idea, we developed two products: one is called Agent Mira™, and the other is called SciClaw, targeting different users. Agent Mira™ is more targeted at enterprises; SciClaw is for professional individual users to help them improve their scientific research operation efficiency.

In the past two years, we have accumulated many customers. In the fields of AI computing power infrastructure, new energy, and beauty R & D, they have achieved a lot of results through our products (including Agent and specific models).

At the same time, we have basically established a business closed - loop. At the beginning, we might just help customers solve a specific small problem, which was a project - service model. Now, we have gradually shifted to an enterprise - oriented deployment subscription model and a potential value - trust cooperation model based on the results achieved by using our tools. So the company's commercialization progress has been very smooth in the past two years.

The entire process is a continuous iteration of frontier research, technology development, and commercialization.

Sell services or products? How can AI4S startups achieve commercialization?

Su Rui @ Yiru Bio

Actually, this is also intertwined with what we mentioned about AI earlier. There are two commercial paths: one is to sell services. For example, some early AI biotech companies mainly focused on services; the other is to sell products. We always believe that selling products is better because if something is truly valuable, it is more important to establish a complete business closed - loop.

But in this process, we found a major problem: there are too many industrial links involved. From biological cultivation, genetic modification, to determining fermentation parameters, to industrial - scale expansion, and then to the subsequent leather products, including injection molding, blow molding and other processes, the complexity is extremely high.

So we chose to reduce the complexity. The core logic is that we don't care about the leather manufacturing process. Others can do it however they like. We just focus on the raw materials.

Looking at it today, this step has played a huge role. Recently, the orders have skyrocketed. An important reason is due to the influence of the geopolitical situation. The crude oil supply in Taiwan, China, Japan and other places has encountered problems. The prices of a large number of plastic products, especially high - grade rubber elastomers as substitutes, have soared. What's more serious is the supply disruption. It's not just about the high price; it's about not being able to buy them at all.

In this context, after we produce the raw materials, we can smoothly enter the market and directly supply them to customers in Taiwan, China, and Japan, which can be used for production immediately. The focus is higher, and the supply flexibility is also stronger.

Speaking of the commercialization logic of AI, we are actually considering two paths now. One is to use AI capabilities within the enterprise to continuously discover new things. For example, our R & D supervisor recently used AI to discover a new molecular metabolic pathway, which is very valuable. But our elastomer market has not reached its upper limit yet, so it is not suitable to invest immediately.

So, can we package the entire "process package" and sell the pipeline to other companies? But the problem is to whom to sell, because there are not many players in the domestic biomanufacturing field who can take over.

From another perspective, is it possible for biomanufacturing to adopt a business model similar to that of new drugs, and license and transfer the process package as an asset? This is a direction that I am still exploring and thinking about.

Zhang Luyang @ Deep Principle

I think our idea has been very simple from the beginning, which is to keep moving forward, continuously address real problems, solve problems when we encounter them, and not get too entangled in the process itself.

Our company has set two directions: First, unlock new materials; second, the unlocked materials must be industrializable. At the same time, our technical route is also based on machine learning. In the early screening stage, we have already considered the possibility of subsequent industrial scale - up, which is in line with the problems we want to solve.

In the process of serving different customers, we also realized that customers at different stages have different needs. Some need a result, while others need a "tool."

We also hesitated at that time. Were we a "tool" company or a company selling material pipelines? Finally, we decided not to limit ourselves according to the traditional R & D and production division of labor. As long as we can solve the specific problems of customers at different stages, it is a good business model.

Therefore, Deep Principle is now a company that "not only conducts product R & D itself, but also formulates material recipes in certain specific fields."

These two businesses form a positive cycle. If we don't even want to use the platform we developed ourselves, it means the product is not good. If we can achieve good material results using our own platform, it naturally proves that the product is good. In this model, we have formed strong momentum.

Our method is to set a big goal first, and at the same time use continuous small goals as signals to make continuous choices and corrections.

What are our advantages in doing AI4S in China?

Su Rui @ Yiru Bio

It has been 5 years since we started our campus - based business. We have completed a relatively complete industrial closed - loop, with revenue, orders, and scale, and it is still growing. This is the foundation brought by our first product.

For us, AI is a continuous optimization and superposition on the basis of China's manufacturing industry. This also explains why so many American companies failed to produce bio - based leather and bio - based elastomers, and ultimately it is our Chinese enterprises that will capture this market. It all boils down to the industrialization foundation.

So I believe that in the long run, the most valuable scenarios for AI for Science in China must be in China's strongest industrial vertical fields.

For example, in the fermentation field where we are located, China accounts for about 70% of the global production capacity; and its leather production accounts for nearly 80% of the world's total. In these two strong fields, AI4S has great potential to further widen the gap with other regions. There is no doubt that China has a very good environment.

Second, in response to the previous mention of the cycle problem: In the years since I started my business, the metaverse and Web3 were popular at that time. Looking back now, many things have changed.

I always tell our investors not to be impatient. Compared with other companies, our greatest advantage is that our first product is already in operation, and we can slowly promote the second product to withstand the cycle. In China, the valuation doubling speed in most industries may not be as extreme as in Silicon Valley. Instead, it allows you to calm down, find real scenarios and real needs, and create real value.

Finally, everyone has to answer the same question: How to create real commercial value? This is the underlying logic.

So we always ask the team several questions: How much efficiency has AI actually improved? What new things has it actually discovered? How should we conduct value assessment? These are the things we continuously discuss.

Zhang Luyang @ Deep Principle

I'll talk from two aspects. I always believe that one should integrate knowledge with action.

First, people are the most important. China has the best scientists and the best engineers. I worked in the United States for a long time. One day, I realized: Why stay in a place that doesn't welcome us and be unhappy at work? When I came back to China, I enjoy my work very much. When people are happy, they will do things better.

Second, and more importantly, AI4S can form a closed - loop in China. It's not as simple as "there are enterprises willing to buy your products."

For example, if we develop a coolant formula, of course, we can use machine learning for screening and start production after setting the specifications. But our starting point is different. In addition to the laboratory side, we will also run the specifications through a series of server simulations. At the same time, we will collaborate with our downstream partners, including data centers and chip companies, and obtain the planning data of the final application scenario before production